Technical Summary: Quality-guided UAV Surface Exploration for 3D Reconstruction

Problem Statement

Autonomous 3D reconstruction and environmental exploration are fundamental tasks for Unmanned Aerial Vehicles (UAVs) in applications ranging from infrastructure inspection to search-and-rescue. However, existing planning strategies often treat exploration (rapid coverage) and reconstruction (high-accuracy mapping) as a single problem, despite their conflicting requirements. Rapid information gathering typically necessitates long-range, fast traversal, while high-fidelity reconstruction requires close-range, precise observations. Furthermore, small UAVs operate under strict computational constraints, requiring lightweight algorithms that can function in real-time on large-scale maps without prior knowledge of the environment's geometry. The core challenge addressed in this paper is developing a Next-Best-View (NBV) planning framework that explicitly adapts to user-defined quality objectives, balancing speed and accuracy dynamically.

Methodology

The proposed approach introduces a modular NBV planning framework that utilizes a Truncated Signed Distance Field (TSDF) to represent the environment. This representation allows the system to encode uncertainty and confidence levels directly into the map structure.

1. Quality Objective Formulation

The authors define a reconstruction quality metric, $Z(M)$, which serves as the optimization target. This metric is derived from the TSDF weights ($w_i$), which represent the confidence in each voxel's occupancy.

• Confidence Modeling: The system leverages the relationship between sensor depth ($d$) and confidence ($w \propto 1/d^2$).
• Target Distance ($d^*$): To achieve a user-defined confidence goal ($w^*$), the system calculates a reference observation distance $d^* = \min(\sqrt{1/w^*}, r_{max})$.
• Quality Function: A modified saturation function $Z_{sat}(M; w^*)$ is used to evaluate the map. This function penalizes areas where confidence exceeds the target (preventing overestimation) and ensures that the average confidence of observed surface voxels meets the predefined threshold.

2. View Generation

Unlike random sampling methods, view candidates are generated based on Surface Frontiers (Incomplete Surface Elements or ISEs)—voxels lying on the boundary between known empty space and unknown space.

• Adaptive Sampling: For each frontier, the system samples three viewing distances: close ($< d^* - \eta$), optimal ($d^* \pm \eta$), and far ($> d^* + \eta$).
• Safety and Alignment: The system checks for collisions and sensor Field-of-View (FoV) constraints. If a direct line of sight is blocked or the angle is incompatible, the viewing direction is rotated horizontally or vertically to find a valid pose while maintaining the target distance.
• Gradient Calculation: Surface normals are computed using a neighbor-weighted gradient method that incorporates uncertainty, ensuring robustness in the presence of unknown voxels.

3. View Evaluation and Selection

The selection of the Next-Best-View ($v^*$) maximizes a combined cost function $J(v) = \alpha J_I(v) + \beta J_N(v)$:

• Information Gain ($J_I$): This component prioritizes frontiers visible from the candidate view. It introduces a visibility probability function that accounts for occlusions by unknown voxels. Crucially, it applies a weighting function $h_v(f)$ that heavily favors views observing frontiers at the optimal distance $d^*$, while penalizing views that are too close (limited field of view) or too far (low confidence).
• Navigation Cost ($J_N$): This component minimizes travel distance and penalizes sharp direction changes relative to the robot's current velocity vector, ensuring smooth and efficient trajectories.

4. Local Planning

The global planner outputs a sequence of poses. A local planner (based on [23]) generates dynamically feasible trajectories, smoothing paths to respect velocity and acceleration constraints. The system also integrates a "freespace pointcloud" to distinguish between empty and unknown space, enhancing the quality of the ESDF (Euclidean Signed Distance Field) used for collision avoidance.

Key Contributions

The paper presents three primary contributions:

1. A New Formalism: A representation-based formalism using TSDF weights to explicitly reflect reconstruction quality and map completeness, allowing for online evaluation of partial maps.
2. Adaptive NBV Strategy: A novel Next-Best-View approach where both view generation and selection stages are driven by the user-defined confidence objective ($Z^*$). This allows the robot to prioritize speed (low $Z^*$) or accuracy (high $Z^*$) dynamically.
3. Extensive Evaluation: A comprehensive evaluation in realistic indoor and outdoor scenarios demonstrating the method's competitiveness in coverage, speed, and map accuracy compared to state-of-the-art baselines.

Experimental Results

The method was validated in ROS1-Gazebo simulations using a quadrotor UAV with a 3D LiDAR in two distinct environments: a complex "House" (open field with scattered objects) and a "Tunnel" (large room with narrow, branching corridors). The performance was compared against an informative sampling method (Online-IPP) and two heuristic baselines (frontier counting and closest frontier).

• Coverage and Efficiency: The proposed method consistently outperformed baselines in coverage ratio. In the "Tunnel" environment, it covered 86% more surface voxels than the second-best approach. It also demonstrated superior path efficiency, with trajectories 3 to 4 times shorter than Online-IPP.
• Accuracy: The method achieved lower mean map errors compared to other strategies, particularly when tuned for high-fidelity reconstruction.
• Adaptability: The system successfully adjusted its behavior based on the target observation distance.
  • High-range targets: Resulted in faster coverage but lower fidelity.
  • Low-range targets: Resulted in slower exploration but significantly higher accuracy (errors 2–4 times lower than other methods in the "House" scenario).
• Consistency: Unlike Online-IPP, which showed high variability and struggled to enter narrow corridors in some trials, the proposed method maintained consistent performance across all trials.

Significance and Claims

The authors claim that their work addresses a critical gap in autonomous robotics: the lack of planning strategies that explicitly account for the trade-off between exploration speed and reconstruction quality. By introducing a quality-guided NBV framework, the paper demonstrates that it is possible to tailor exploration behavior to specific mission goals (e.g., rapid survey vs. precise inspection) without sacrificing overall efficiency.

The paper asserts that this is the first NBV approach to adapt both view generation and selection to a required confidence level on the final map. The results suggest that incorporating explicit quality criteria into planning algorithms yields superior performance in terms of coverage, path efficiency, and final map accuracy compared to conventional strategies that treat these objectives as secondary or implicit. The authors conclude that this approach enables more efficient, task-driven exploration for aerial robots in unknown environments.